NISQA Speech Quality Corpus

NISQA 语料库包括 14,000 多个模拟（例如编解码器、丢包、背景噪声）和实时（例如手机、Zoom、Skype、WhatsApp）的语音样本。每个文件都标有对整体质量的主观评级以及质量维度的噪声、着色、不连续性和响度。总的来说，它包含超过 97,000 个人对每个维度和整体 MOS 的评分。 NISQA 语音质量语料库包含两个训练、两个验证和四个测试数据集：NISQA_TRAIN_SIM 和 NISQA_VAL_SIM：包含来自四个不同数据集的语音样本的模拟失真。分为训练集和验证集。 NISQA_TRAIN_LIVE 和 NISQA_VAL_LIVE：包含带有 Lirivox 有声读物样本的实时电话和 Skype 录音。分为训练集和验证集。 NISQA_TEST_LIVETALK：包含真实电话和 VoIP 通话的录音。 NISQA_TEST_FOR：包含来自法医语音数据集的实时和模拟语音样本。 NISQA_TEST_NSC：包含来自 NSC 数据集的实时和模拟语音样本。 NISQA_TEST_P501：包含来自 ITU-T Rec. 的实时和模拟语音样本。第 501 页。数据集是根据使用的源语音和噪声样本的原始条款提供的。有关数据集和许可证的更多详细信息，请参阅 NISQA_Corpus.zip 中每个数据集文件夹中的单独自述文件和许可证文件。一般来说，这个语料库中的所有文件都可以用于非商业研究目的，一些数据集也可以用于商业目的。

The NISQA corpus comprises over 14,000 speech samples with both simulated distortions (e.g., codecs, packet loss, background noise) and real-world recordings (e.g., mobile phone, Zoom, Skype, WhatsApp). Each sample is annotated with subjective ratings for overall quality, as well as four quality dimensions: noise, coloration, discontinuity, and loudness. In total, the corpus contains over 97,000 human ratings for each dimension and the overall Mean Opinion Score (MOS).

The NISQA speech quality corpus includes two training, two validation, and four test datasets:
1. NISQA_TRAIN_SIM and NISQA_VAL_SIM: These datasets contain speech samples with simulated distortions derived from four distinct datasets, split into training and validation subsets respectively.
2. NISQA_TRAIN_LIVE and NISQA_VAL_LIVE: These datasets consist of real-world telephone and Skype recordings paired with Lirivox audiobook samples, split into training and validation subsets respectively.
3. NISQA_TEST_LIVETALK: Contains recordings of real telephone and VoIP calls.
4. NISQA_TEST_FOR: Includes real and simulated speech samples sourced from a forensic speech dataset.
5. NISQA_TEST_NSC: Includes real and simulated speech samples from the NSC dataset.
6. NISQA_TEST_P501: Includes real and simulated speech samples from ITU-T Recommendation P.501.

All datasets are provided under the original terms of the source speech and noise samples used. For more details about the datasets and their respective licenses, please refer to the individual README and license files contained within each dataset folder in the NISQA_Corpus.zip archive. Generally, all files in this corpus may be used for non-commercial research purposes, and some datasets are also permitted for commercial use.

搜集汇总

背景与挑战

背景概述

NISQA Speech Quality Corpus是一个用于语音质量评估的数据集，包含超过14,000个模拟和实时语音样本，每个样本都标有整体质量和多个维度的主观评分，总计超过97,000个人类评分。数据集分为多个训练、验证和测试子集，适用于非商业研究目的，由德国人工智能研究中心于2021年发布，主要用于支持语音质量预测模型的研究。